Klarna’s AI Agents Cut Support Costs by 40%, Drive $60M Savings

Klarna's AI Agents Cut Support Costs by 40%, Drive $60M Savings
Key Takeaways

  • Klarna’s multi-agent customer service system handled the equivalent workload of 853 full-time employees, cutting cost per transaction 40% and delivering $60 million in savings by Q3 2025.
  • In financial services, multi-agent architectures have cut risk analysis latency to 0.6 seconds by parallelising queries across credit bureau APIs, transaction databases and regulatory lists, while reducing false positives by 34% through multi-signal correlation.
  • Orchestration costs can scale sharply: a task costing $0.10 in API calls for a single agent may cost $1.50 in a multi-agent system, and operational costs typically represent the majority of three-year total spending, far exceeding initial build costs of $75,000 to $500,000.

Klarna‘s AI customer service deployment now handles the workload of 853 full-time employees across 23 markets, at roughly 60 cents on the dollar compared to what human agents cost. That single data point captures why enterprise interest in multi-agent AI has moved from experimental to urgent, and why the deployment challenges are proving harder than the demos suggested.

Enterprise Multi-Agent AI Surges

Gartner recorded a 1,445% increase in multi-agent system inquiries between Q1 2024 and Q2 2025. That number reflects a shift in what enterprises actually want from AI: not better text generation, but systems that can research, plan, coordinate tasks, use tools and execute work across complex operational environments. Single-agent deployments, even well-tuned ones, hit a ceiling on multi-step workflows. Multi-agent architectures are how organisations are trying to push past it.

The underlying driver is operational execution. Earlier AI adoption focused on improving response quality through techniques like retrieval-augmented generation (RAG). Current enterprise expectations have moved further: systems must decompose ambiguous problems, delegate to specialised agents, validate outputs and complete work end-to-end. Multi-agent systems attempt to mirror how human teams operate, each agent handling a distinct function, with coordination handled by an orchestration layer above them.

Proven Cost Savings and Efficiency Gains

Klarna’s deployment is the most cited case study in the space. The system manages approximately two-thirds of all customer inquiries across 23 markets in more than 35 languages, around the clock. Response times improved by 82%, cost per transaction fell 40% from $0.32 in Q1 2023 to $0.19 in Q1 2025, and repeat contacts dropped by 25%. The $60 million in savings by Q3 2025 came primarily from that cost-per-transaction compression at scale. Notably, these figures came after Klarna reversed an earlier full-automation push: in May 2025, the company began rehiring human agents following complaints about response quality on complex cases, with CEO Sebastian Siemiatkowski acknowledging the AI-only approach had produced ‘lower quality’ outcomes and committing to guarantee customers access to a human when needed.

Other deployments point in the same direction. Doxy.me, a telemedicine platform, shifted a significant share of inbound calls from a legacy IVR system to an AI Voice Agent, with the agent handling more than 30% of calls compared to 5% under the prior system. An unnamed e-commerce platform managing more than 50,000 daily interactions reported resolution times falling by 58%, first-call resolution reaching 84% and operating costs dropping by 45% after deploying a multi-agent customer service system. These figures are drawn from vendor and operator disclosures rather than independent audits, and real-world results will vary by deployment quality.

Complex Problem Solving in Action

Financial services is where multi-agent architectures are getting their most demanding tests. A risk analysis deployment described in industry coverage achieved 74% faster processing by parallelising queries across credit bureau APIs, internal transaction databases and regulatory lists simultaneously, cutting total latency to 0.6 seconds. The same architecture improved risk detection accuracy by 23-27% and cut false positives by 34% through multi-signal correlation.

JPMorgan Chase‘s “Ask David” investment research system is reported to have improved accuracy from 50% to over 90% through iterative development and evaluation-driven refinement. Mayo Clinic’s AI agents have demonstrated 89% diagnostic accuracy in complex cases and cut diagnostic time by 60%, according to reports. In supply chain, General Mills projects more than $20 million in supply chain savings and over $50 million in waste reduction from its agentic AI programmes, coordinating agents across inventory planning, demand forecasting, route optimisation and supplier monitoring.

Software development is a different kind of test case. MetaGPT is a framework that simulates an entire software engineering team, with specialised agents acting as product managers, architects, developers and QA engineers to complete projects end-to-end. The practical question, how reliably it handles the handoffs between those roles on real codebases, remains harder to verify from public material.

Navigating Deployment Challenges

Non-determinism is the production problem that most enterprise teams underestimate. LLM-based agents can produce unpredictable outputs, compound errors across steps, or enter loops that are difficult to detect and break. In regulated industries like finance and healthcare, that variability is a compliance problem, not just a quality one. Teams that have successfully moved to production tend to constrain agent behaviour explicitly: forcing structured outputs, restricting agents to predefined action sets rather than open-ended reasoning, and building hard stops into orchestration logic.

Legacy integration is the other major friction point. Agent demos typically assume clean APIs and structured data. Production environments have neither: data sits in older databases, proprietary formats and SOAP services that agents cannot access directly. Successful deployments build a dedicated tooling layer, a set of standardised internal APIs that agents call safely, rather than allowing direct access to underlying systems. That layer adds build time and maintenance cost, but it is what makes orchestration predictable.

Cost scaling is the challenge that catches enterprises off guard most often. A task costing $0.10 in API calls for a single agent can reach $1.50 in a multi-agent system, driven by the orchestration overhead of maintaining context across agents and the cross-agent verification calls that compound with each additional layer. Operational costs typically represent the majority of three-year total spend, well above initial build costs that typically range from $75,000 to $500,000 for enterprise deployments. Teams that scope only for build cost tend to find the economics deteriorate quickly at scale. Governance frameworks, continuous testing and transparent cost monitoring are not optional additions, they are what keep multi-agent systems financially viable in production. For more on the economics of agentic AI at scale, see our coverage of why unlimited agentic AI subscriptions are unsustainable.

Maturing Frameworks and What Comes Next

The tooling layer for multi-agent systems has matured considerably. LangGraph handles stateful agent graphs with human-in-the-loop checkpoints. CrewAI offers role-based agent coordination with a lower barrier to entry. Microsoft‘s AutoGen has moved from a research framework to a production-grade option for Azure-native teams, with an event-driven architecture that supports distributed deployments. AWS’s Multi-Agent Orchestrator integrates directly with Bedrock and Lambda for teams already in that stack. The recent updates to CrewAI and LlamaIndex have also pushed agentic research costs down sharply, which changes the calculus for teams evaluating build costs.

None of these frameworks solves the hard problems automatically. Modularity, communication protocol design and the separation of orchestration from execution still require deliberate architectural choices. Human-in-the-loop controls remain important wherever agent errors carry regulatory or financial consequences. What the maturing framework landscape does offer is a shorter path from proof-of-concept to production, provided teams invest in the governance and monitoring infrastructure alongside the agents themselves. For more analysis on enterprise AI strategy, visit our Enterprise AI section.

Morgan Blake
Morgan Blake

Morgan is a technology analyst covering enterprise AI strategy, automation, and business transformation. Morgan tracks how organisations are deploying AI at scale.

📰 Journalists welcome — cite Auton AI News with attribution. Press & Media → | press@autonainews.com